Re: [PATCH v4 05/11] mm/migrate: clear __GFP_RECLAIM for THP allocation for migration

[Joonsoo Kim]

On Tue, Jul 07, 2020 at 02:17:55PM +0200, Vlastimil Babka wrote:
> On 7/7/20 9:44 AM, js1304@xxxxxxxxx wrote:
> > From: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
> > 
> > In mm/migrate.c, THP allocation for migration is called with the provided
> > gfp_mask | GFP_TRANSHUGE. This gfp_mask contains __GFP_RECLAIM and it
> > would be conflict with the intention of the GFP_TRANSHUGE.
> > 
> > GFP_TRANSHUGE/GFP_TRANSHUGE_LIGHT is introduced to control the reclaim
> > behaviour by well defined manner since overhead of THP allocation is
> > quite large and the whole system could suffer from it. So, they deals
> > with __GFP_RECLAIM mask deliberately. If gfp_mask contains __GFP_RECLAIM
> > and uses gfp_mask | GFP_TRANSHUGE(_LIGHT) for THP allocation, it means
> > that it breaks the purpose of the GFP_TRANSHUGE(_LIGHT).
> > 
> > This patch fixes this situation by clearing __GFP_RECLAIM in provided
> > gfp_mask. Note that there are some other THP allocations for migration
> > and they just uses GFP_TRANSHUGE(_LIGHT) directly. This patch would make
> > all THP allocation for migration consistent.
> > 
> > Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
> > ---
> >  mm/migrate.c | 5 +++++
> >  1 file changed, 5 insertions(+)
> > 
> > diff --git a/mm/migrate.c b/mm/migrate.c
> > index 02b31fe..ecd7615 100644
> > --- a/mm/migrate.c
> > +++ b/mm/migrate.c
> > @@ -1547,6 +1547,11 @@ struct page *new_page_nodemask(struct page *page,
> >  	}
> >  
> >  	if (PageTransHuge(page)) {
> > +		/*
> > +		 * clear __GFP_RECALIM since GFP_TRANSHUGE is the gfp_mask
> > +		 * that chooses the reclaim masks deliberately.
> > +		 */
> > +		gfp_mask &= ~__GFP_RECLAIM;
> >  		gfp_mask |= GFP_TRANSHUGE;
> 
> In addition to what Michal said...
> 
> The mask is not passed to this function, so I would just redefine it, as is done
> in the hugetlb case. We probably don't even need the __GFP_RETRY_MAYFAIL for the
> THP case asi it's just there to prevent OOM kill (per commit 0f55685627d6d ) and
> the costly order of THP is enough for that.

Will check.

Thanks.